The pervasiveness of plastics in modern life stems from their diverse chemistries and tunable properties, offering flexibility, strength, and conductivity at low cost and weight. However, the enormous global production of plastics (368 million tonnes in 2019) and the short lifespans of many plastic products, especially packaging, lead to significant environmental pollution. Packaging plastics represent a large source of pollution in ecosystems, posing a serious threat due to their persistence as waste and microplastics. The search for eco-friendly alternatives, specifically bioplastics with properties similar to conventional plastics but offering sustainable recycling options, is crucial for a circular economy. Polyhydroxyalkanoates (PHAs) are a promising class of bio-derived and biodegradable materials, synthesizable by microorganisms using sunlight and CO2. Their diverse chemistries and tunable properties, controlled by parameters such as the number of carbons in the main chain and side chain, and the terminating functional groups of the side chain, offer a wide range of mechanical and thermal properties. Copolymers further expand this property space, allowing for the combination of multiple PHA motifs or PHAs with conventional polymers. The immense number of possible copolymer compositions (over a million for PHA-only copolymers alone), however, makes traditional high-throughput experimentation impractical. Polymer informatics, using data-driven machine learning, offers a powerful alternative approach to search for application-specific candidate materials. This approach avoids the limitations of resource-intensive, time-consuming methods like density functional theory (DFT) or classical molecular dynamics (MD) simulations.

Existing literature extensively documents the properties and potential applications of PHAs. Studies have explored their biosynthesis, processing, and industrialization, highlighting the tunability of mechanical and thermal properties such as Young's modulus, tensile strength, elongation, glass transition temperature, melting temperature, and degradation temperature. Previous work demonstrated the impact of structural modifications, copolymerization, and blending on PHA properties. Specifically, increasing the number of carbon atoms in the backbone increases elongation and mechanical strength while enhancing degradability. Side-chain-terminating phenyl groups increase glass transition temperatures due to enhanced inter-chain interactions. Copolymers of PHAs have shown improvements in mechanical properties while maintaining desirable temperature operating windows. Blending PHAs with conventional polymers can create synergistic effects, leading to recyclable materials with enhanced properties and gas permeability control, crucial for food packaging. While some studies have employed machine learning for predicting individual PHA properties, a comprehensive multitask approach predicting a wide array of properties for both homo- and copolymers remains relatively unexplored. Polymer informatics has emerged as a valuable tool in materials discovery and design, but its application to the vast chemical space of PHAs and their copolymers has significant potential.

This study utilizes a multitask deep neural network approach to predict polymer properties. The dataset comprised 22,731 data points (approximately 60% homopolymers, 30% copolymers) covering thermal (*Tg*, *Tm*, *Td*), mechanical (*E*, *σy*, *σb*, *εb*), and gas permeability (*μO2*, *μCO2*, *μN2*, *μH2*, *μHe*, *μCH4*) properties. Data curation involved averaging multiple data points for a single polymer, selecting only data points measured using consistent methods (e.g., DSC for *Tg* and *Tm*, TGA for *Td*), and utilizing a DBSCAN clustering algorithm to identify and address outliers. All property values were scaled to the range [0, 1] before training. Gas permeabilities were log-transformed due to their power-law distributions. Three multitask deep neural networks, one for each property category (thermal, mechanical, gas permeability), were trained. The architecture involved three separate predictors, one for each selector vector (S1, S2, S3), a binary vector selecting the output property. A meta-learner was used to combine predictions from these predictors. The bioplastic search space was defined using 540 PHAs (variations in carbon atoms and side-chain functional groups) and 13 conventional polymers, generating approximately 1.4 million bioplastic candidates (including homopolymers, PHA-only copolymers, and PHA-conventional polymer copolymers). A two-step protocol was used to identify bio-replacements for seven commonly used petroleum-based plastics: a nearest-neighbor search to identify the five closest replacements in each copolymer subgroup (PHA-only and PHA-conventional), followed by expert selection based on synthesizability. Uniform Manifold Approximation and Projection (UMAP) was used to visualize the polymer fingerprints in the search space.

The developed multitask deep neural networks demonstrated exceptional performance, with high R² values (0.97 for meta-learner and 0.93 for cross-validation). The meta-learner achieved R² values of 0.98, 0.97, and 0.96 for *Tg*, *Tm*, and *Td*, respectively, and high R² values for mechanical and gas permeability properties (0.99 or higher for gas permeability). The UMAP visualization showed clear clustering of polymers with similar chemistries, validating the efficacy of the fingerprints in creating a meaningful learning space. Property relationships within the bioplastic search space confirmed expected physical trends (e.g., correlation between *Tg*, *Tm*, and *E*). The two-step selection protocol identified 14 promising PHA-based bioplastic replacements for seven commodity plastics (PE, PP, PVC, PET, PS, Nylon 6, PEN), accounting for over 75% of annual plastic production in Europe. These replacements showed property profiles closely matching the target plastics in radar charts. The identified bio-replacements predominantly contained aromatic groups in their side-chains. The authors discuss potential synthesis routes, including biosynthesis (leveraging existing work on aromatic-containing PHAs and the potential of CRISPRi for further advancements) and chemical synthesis (using methods similar to those employed for polystyrene-co-lactone, PHA/PEO, and phthalate-co-lactones copolymers).

This study successfully demonstrated the utility of multitask deep neural networks for predicting the properties of a vast range of bioplastics and identifying promising alternatives to conventional plastics. The high accuracy of the prediction models and the comprehensive search space allowed for the identification of suitable bio-replacements that closely mimic the properties of widely used petroleum-based plastics. The feasibility of synthesizing the identified bioplastics, through both bio- and chemical routes, significantly enhances the translational potential of this research. The findings highlight the power of polymer informatics in accelerating the discovery and design of sustainable materials, thereby contributing to the development of a circular economy and mitigating the environmental impact of plastic waste. The relatively close match of properties between the identified bio-replacements and the target plastics suggests a viable path towards replacing conventional plastics with more environmentally sustainable alternatives, potentially leading to superior performance and reducing reliance on petroleum resources.

This work presents a novel informatics-based bioplastic design pipeline, successfully identifying promising PHA-based replacements for seven major petroleum-based commodity plastics. Multitask deep neural networks achieved high accuracy in predicting multiple polymer properties, facilitating a comprehensive search within a massive bioplastic candidate space. The identified bio-replacements exhibit properties closely resembling those of conventional plastics, opening doors for future research in optimizing synthesis routes and scalability. This approach offers a significant advancement in accelerating the transition to sustainable materials and reducing our dependence on non-renewable resources. Future work could focus on refining the prediction models by incorporating factors such as molecular weight distribution, morphology details, and processing parameters. Further investigation into scalable and cost-effective synthesis routes for the identified bioplastics is warranted.

The current study has some limitations. The property predictors do not fully account for process and manufacturing conditions, morphology details (like crystallinity), or the influence of additives. The molecular weight distribution, chain branching, and subtle configurational variations are also not explicitly considered. While the data curation efforts were extensive, potential biases or inaccuracies in the underlying experimental data could affect the model's accuracy. The synthesizability assessment is qualitative and relies on existing literature; a more quantitative analysis integrating reaction yield and cost would be beneficial. Finally, the study's focus on specific commodity plastics may limit the generalizability of findings to other applications and plastic types.